Computer-assisted language mediation in teaching human-centred augmented translation

Chapter 8 Computer-assisted language mediation in teaching human-centred augmented translation Maria Zimina-Poirot Université Paris Cité, France In the evolving landscape of translation education, the integration of computerassisted language mediation has emerged as a key strategy for teaching humancentred augmented translation. This study explores the implementation of generative AI (GenAI) models, in particular Large Language Models (LLMs), in translation workflows to enhance translator competence while maintaining ethical standards. Using custom AI systems that adapt and evolve with new translations, students can be trained to critically evaluate AI-generated content and manage hybrid translation workflows that combine AI output with human expertise. The research emphasises the importance of understanding the limitations and potential biases of AI, and advocates a balanced approach in which AI augments rather than replaces human judgement. Practical examples and exercises demonstrate the strengths and weaknesses of using AI in information processing and multilingual text generation, with the ultimate aim of training a new generation of translators to use AI technologies responsibly. 1 Introduction 1.1 GenAI in translation education In translation education, integrating generative AI (GenAI) models into translation processes presents challenges in maintaining translator autonomy while embracing technology (O’Brien 2024). Ethical concerns, including privacy, algorithmic bias, and responsible use of machine translation and linguistic data, Maria Zimina-Poirot. 2026. Computer-assisted language mediation in teaching human-centred augmented translation. In JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, 147–165. Berlin: Language Science Press. DOI: 10.5281/zenodo.17641078 Maria Zimina-Poirot must be carefully considered as translation education undergoes a paradigm shift (Ramírez-Polo & Vargas-Sierra 2023, Slimi & Carballido 2023, Tavares et al. 2023). Developing computer-assisted language mediation in translation education requires successfully integrating data analysis and AI assistants into pedagogical workflows. In this respect, AI is one of the building blocks of translation technologies, along with computer-assisted translation environments, project management software, ontologies, expert involvement and data management facilities (Garcia 2023, He et al. 2023, Mitchell-Schuitevoerder 2020). 1.2 Computer-assisted language mediation Used as part of a computer-assisted dynamic mediation framework (Hu et al. 2024), GenAI tools allow translation students not only to produce multilingual content, but also to compare and fine-tune the generated output, fostering a deeper contextual understanding of the strengths and weaknesses of different translation models and enabling future translation professionals to become more proficient reviewers of automatically generated content (Zhuang et al. 2024). This process requires new translation skills and up-to-date knowledge of AI assistants built on underlying AI models, such as Large Language Models (LLMs) (Douglas 2023). As LLM technology is introduced to the field of translation, enabling trainees to navigate language variations within adaptive pathways, the concept of a linguistic norm emerges as an increasingly crucial analytical tool, facilitating efficient translation across diverse contexts (Fabricius 2022, Schulz & Ollig 2023, Sinner 2020). From this perspective, computer-assisted language mediation can be taught as a set of new translation competences, including the ability to create and use a language assistant, and potentially to understand how to employ LLM-enhanced linguistic data responsibly (Raza et al. 2025). From a technical standpoint, building personalised AI assistants from scratch requires substantial resources, emphasising the efficiency of leveraging pretrained models (Shichkina & Krinkin 2022). This process can be part of training programs specifically tailored to prepare data scientists specialised in translation. Few-shot learning (FSL),1particularly in-context learning (ICL),2empowers 1Few-Shot Learning (FSL) is a machine learning approach that enables models to adapt to new tasks with only a small amount of labelled data (Mosbach et al. 2023). For more information: https://medium.com/ubiai-nlp/step-by-step-guide-to-mastering-few-shotlearning-a673054167a0 2In-context learning, as popularised in LLMs such as GPT-3 and GPT-4, involves a model’s capacity to comprehend and execute tasks based on contextual information embedded within the input sequence, without modifying the model’s parameters (Dong et al. 2024). 148 8 Computer-assisted LM in teaching human-centred AT translators with limited natural language processing (NLP) expertise. This approach allows LLMs to rapidly acquire new capabilities, thereby facilitating seamless integration of AI into a range of translation workflows. In this regard, prompt engineering3facilitates the development of a versatile approach that can be applied to a range of tasks, including multilingual translation, fine-tuning of translation models (Doğru & Moorkens 2024), information extraction, and text summarisation. Training future translators to effectively prompt LLMs for these tasks in digital communication is essential. Our analysis in Section 3 will provide specific examples that illustrate these approaches in teaching. 1.3 Investigating biases and inconsistencies of GenAI An important area of research in teaching GenAI technologies in translation courses is to make students aware of the potential biases and inconsistencies they may encounter when using this new generation of tools. The shortcomings of AIgenerated content become apparent when the translation task is fully mastered. Unfortunately, if students are not fully proficient in the target language and lack specialised knowledge and translation practice, they may find it difficult to effectively assess the quality of the output (Tavares et al. 2023). Understanding both the strengths and limitations of GenAI through a series of practical examples is vital for efficient use of translation tools (Farrell 2023). By exposing students to real-life scenarios and practical exercises, it is possible to develop a better understanding of the qualities and shortcomings of automatically generated content in the context of multilingual communication. This hands-on approach allows trainee translators to identify potential biases, inconsistencies and errors that may arise from relying solely on AI assistance. In addition, by analysing and discussing these practical examples, students can learn to critically evaluate the output of AI assistants and identify when human intervention or additional expertise is required to produce accurate and contextually relevant content. This knowledge is also essential for developing a new generation of translation-revision workflows that successfully combine GenAI and human expertise. 3The field of prompt engineering is concerned with the design and optimisation of prompts to help a model perform better on new tasks with limited data. For prompts to be effective, they must clearly define the task at hand, include context to help the model use its existing knowledge, and gradually introduce new information or tasks. 149 Maria Zimina-Poirot 2 Contextual understanding and automation Artificial intelligence is now capable of processing a wide variety of information, including text, images, speech, facial expressions, and, to a certain extent, body movements. These capabilities can be used to develop a large panel of AI assistants (Yang et al. 2024). One of the most intriguing aspects of these technologies is their handling of context, which differs significantly from human understanding (O’Brien 2024, Zhu et al. 2024). Human understanding involvesa combination of experience, emotion, cultural background and situational awareness. This allows people to naturally interpret nuances, implicit meanings and subtleties in communication. Assessing human understanding of context through written expression has long been a focus of research, with methodologies evolving to encompass cognitive, linguistic, and technological perspectives (O’Brien 2024). Researchers in cognitive psychology and linguistics have devoted considerable time and effort to studying the mechanisms of language processing, with a particular focus on the role of lexical access, syntactic parsing and discourse coherence in comprehension. In addition, discourse analysis has provided valuable insights into how contextual cues shape interpretation, revealing complex interplay between language and cognition in human comprehension (Van Dijk 2006, Pleyer & Winters 2015, Gledhill & Pecman 2018). In contrast, GenAI models process context through the identification of patterns within the data they learned during the training phase (Zhuang et al. 2024). In contrast to a true understanding of context, these models rely on vast quantities of data to identify and generate responses based on statistical correlations. To illustrate, when presented with a sentence, an AI assistant analyses the sequence of words and predicts the most probable continuation based on the patterns it has learned. Currently, this approach can effectively imitate human-like responses, but it lacks genuine comprehension and the capacity to grasp more profound meanings or unexpected cultural nuances. Additionally, it tends to accentuate recognised patterns in the generated output. This difference in contextual understanding has significant implications for translation. While AI may be good at generating coherent and contextually appropriate text, it may struggle with tasks that require cultural sensitivity or complex problem-solving that requires nuanced human judgement. Furthermore, AI’s understanding of context is limited to the data on which it has been trained. If training data sets lack diversity or are biased, AI-generated content will reflect these limitations (Martínez et al. 2023). Understanding these differences is 150 8 Computer-assisted LM in teaching human-centred AT crucial when integrating GenAI assistants into translation workflows. While AIbased technologies can significantly increase productivity and efficiency by performing routine tasks and providing quick information, they should complement rather than replace human judgement and expertise in translating certain types of content (Denning & Arquilla 2022). Balancing the strengths of AI in processing and generating text with advanced human skills used to understand and interpret complex contexts can lead to more efficient and nuanced applications of translation technology for translation project management, known as human-centred augmented translation (O’Brien 2024). In practical terms, what specific tasks can help students understand GenAI weaknesses as reflected in the warning “ChatGPT can make mistakes. Check important info.” displayed by ChatGPT (https://chatgpt.com)? In what types of situations are these tools helpful, and what are the potential pitfalls? The following section explores the point of studying biases and inconsistencies in GenAI operation through practical examples of classroom activities. 3 Exploring AI in practice: delving into knowledge-rich texts 3.1 Text summarisation and interactive learning with AI assistants In this section, we consider how practical tasks can help students gain a deeper understanding of GenAI. One of the easiest ways to address these complex processes in translation education is to use examples of knowledge-rich texts (Gledhill & Kübler 2016). In this context, information synthesis is one of the key capabilities that AI brings to streamline information processing, with tools such as SciSpace and PDFgear being able to summarise key information almost instantly for efficient retrieval.4In this respect, linguistic standards, specifications and style guides provide an excellent basis for practice (Gledhill & ZiminaPoirot 2023). These well-structured documents precisely outline the standards that should be applied in written communication. At the same time, the length and complexity of these documents often require detailed human analysis and extensive practice to fully master the rules. To illustrate how AI can be used in teaching, let us consider a specific example. In a Master’s level 2 course on technical translation and controlled languages, stu4SciSpace is an advanced platform designed for researchers that uses natural language processing and machine learning to facilitate efficient literature review and analysis in scientific fields. 151 Maria Zimina-Poirot dents are expected to familiarise themselves with ASD-STE100 (Simplified Technical English).5ASD-STE100 is an international specification for writing technical documentation in a controlled natural language.6This comprehensive document of over four hundred pages requires considerable effort to master the principles of STE for writing procedures, descriptive technical texts, warnings and cautions. In addition, there is an extensive “core dictionary” containing a wide range of lexical items that students must learn to use in context. The dictionary is accompanied by illustrative examples and contrasting cases (APPROVED: STE/ not approved: non-STE). Integrating AI into ASD-STE teaching can potentially improve the learning experience. The following examples highlight several ways in which this can be achieved. • Use case: Use tools such as PDFGear or SciSpace to quickly summarise lengthy sections of ASD-STE100. • Benefit: By using a concise format in classroom settings with limited teaching time, students can focus on understanding and practical applications of the STE rules rather than being overwhelmed by the volume of text. In this context, AI assistants can act as interactive tutors, answering students’ questions about specific language policies and providing immediate feedback. Prompt clarification and reinforcement can accelerate learning. In addition, AI assistants can quickly locate specific sections or rules within lengthy documents through enhanced research and reference, saving time and improving efficiency. The main benefit is that students can spend more time applying the rules and less time searching for them. At the same time, through this type of exercise, students will realise that both PDFGear and SciSpace copilots are quite effective at extracting relevant knowledge and providing a comprehensive list of essential rules to be applied in writing. Nevertheless, this knowledge extraction is based on patterns in the text and may be superficial when it comes to applying these rules in writing and rewriting 5See curriculum for Master 2 ILTS, Université Paris: https://odf.u-paris.fr/fr/offre-deformation/master-XB/arts-lettres-langues-ALL/traduction-interpretation-K6JMSAFS/ master-traduction-interpretation-parcours-industrie-de-la-langue-et-traduction-specialiseeJRQNFV5Z.html 6ASD-STE100 (https://www.asd-ste100.org) is a set of writing rules and a controlled vocabulary for producing clear and simplified technical documentation in English. It was developed in the 1980s by the aerospace industry associations (AECMA/ASD and AIA) to make aircraft maintenance documentation more readable for people with limited English proficiency. 152 8 Computer-assisted LM in teaching human-centred AT texts. Students can rapidly become aware of these limitations and gain insight into the technical aspects of producing content with GenAI. This dual awareness can help them to understand both the potential and the limitations of using AI assistants in the context of technical communication and translation: • Use case: Use PDFgear and SciSpace copilots to generate examples and counter-examples based on the core dictionary and ASD-STE100 rules, then critically evaluate the AI’s output. • Benefits: Firstly, this type of exercise encourages active learning and critical thinking as students compare AI-generated content with examples of human rewriting. Secondly, it provides insight into the limitations of automatic text generation in the practical application of writing rules. The following section will demonstrate how previous exercises can be further developed into a more conclusive experiment, using tools such as ChatGPT, Perplexity AI, or Claude.ai.7In this case, a prompt language model (input in Technical English and output in Simplified Technical English) would be employed to demonstrate the outcomes of few-shot learning on a specific data set. In addition, prompts can be used to highlight inconsistencies in the generated output, thereby directing attention towards potential enhancements (such as the avoidance of noun clusters and the use of articles encouraged by STE). 3.2 AI output refinement using prompt engineering and few-shot learning From a technical perspective, building LLMs from scratch requires significant resources, which highlights the efficiency of using pre-trained models. This process can be integrated into teaching programs specifically designed to train data scientists specialising in translation. Few-shot learning, especially in-context learning, enables translators with limited NLP expertise to quickly acquire new skills and seamlessly integrate AI into translation workflows across multiple applications. In the context of few-shot learning, a model is presented with a limited number of labelled examples corresponding to a desired input-output scenario. Transfer learning principles are embedded in the model to enable it to use these examples 7ChatGPT is a conversational AI model developed by OpenAI based on the GPT (Generative Pre-trained Transformer) architecture. Based on OpenAI’s model, and a standalone LLM with NLP capabilities, Perplexity AI is a conversational search engine that answers queries using natural language predictive text, using sources from the web. Claude.ai is trained by Anthropic using Constitutional AI (Bai et al. 2022). 153 Maria Zimina-Poirot to adjust its parameters and fine-tune its representations to enhance the presence of specific patterns. The main advantage of few-shot learning is its ability to generalise from limited examples to unseen data, due to the rich abstract representations acquired during pre-training. However, even in this scenario, the results may not be entirely satisfactory and may require human refinement. • Use case: Create a prompt language model with labelled examples of “Input text in Technical English” and “Output text in STE”, demonstrating the results of few-shot learning on a given data sample. • Benefit: This hands-on exercise allows students to interact directly with the technology and gain practical experience using prompt language models for text generation tasks. In addition, this approach provides an opportunity for students to observe the limitations of AI in fully capturing complex linguistic nuances, highlighting the importance of human analysis and editing to refine the output for satisfactory results. 3.3 Terminology extraction with LLMs: understanding model limitations In parallel, other initiatives are exploiting the capabilities of combining AI and NLP to facilitate access to specialised knowledge. Teaching terminology extraction from multimodal sources using LLMs is one of the areas which can reinforce learners’ augmented reasoning in translation. Before using LLMs with multimodal capabilities (such as Perplexity AI playground: https://labs.perplexity.ai or https://gemini.google.com) for the analysis and extraction of terms from multiple sources, it is beneficial for students to contrast the outcomes generated by LLMs with those produced by conventional term extraction, text-mining, and text annotation tools. This comparison enables learners to become aware of potential biases and hallucinations. For example, the frequencies of candidate terms identified by GenAI are not always 100% accurate compared to traditional text segmentation and term extraction methods. AI models can propagate errors from misinterpretations in the text, resulting in inaccurate term frequencies. Misreading sentence structure can lead to incorrect counts of certain terms (Uchida 2024). In addition, AI models trained on diverse and large datasets may inappropriately apply patterns learned in one context to another. For example, if a model 154 8 Computer-assisted LM in teaching human-centred AT has frequently seen “climate change” discussed alongside terms such as “policy” and “regulation”, it might overestimate the frequency of these terms in any document mentioning “climate change”, even if the specific text focuses on scientific data rather than policy. These errors highlight a key limitation of AI models: their reliance on patterns and associations learned from vast amounts of data, which can sometimes lead to misinterpretations if the specific context of a new text differs from those patterns. The generated output exaggerates the learned patterns. As a result, the frequency of terms identified by GenAI may not always accurately reflect the actual text and its context, especially in cases where a nuanced understanding of language is required. Therefore, comparing AI-generated results with those of traditional text segmentation and term extraction tools, which often use straightforward statistical methods, can help identify and correct such inaccuracies. • Use case: Use LLMs to extract terms from different types of documents. Check the extracted terms for accuracy and relevance. Use text-mining and navigation tools (such as Voyant Tools: https://voyant-tools.org) to critically evaluate the results produced by AI assistants. • Benefit: Learn to use quantitative methods to evaluate the quality of AIgenerated text analysis. Note that ready-made prompts for text analysis can be found on many websites.8 Such models can be adapted, for example, for term extraction (works for Perplexity AI): • Use case (sample prompt): Extract the most relevant terms with contexts from the attached document. Present the results in a markdown table with the following columns: term, frequency, contexts, definition.9 3.4 Towards “augmented reasoning” in translation: linguistic data and LLMs Many other experiments with AI tools are possible in the classroom, both to demonstrate the role that AI models can play in different application settings, 8For example, Keywords Everywhere ChatGPT Prompt Templates: https:// keywordseverywhere.com/chatgpt-prompt-templates.html (also available for Claude.ai). 9More examples are available in #2024TEF -AI-powered terminology extraction: A hands-on guide for translators by Josh Goldsmith: https://youtu.be/5Y5PhzyeMGI?list= PLLqIRaiVCGCR80tysPOQ2AHJ30HEyItXX 155 Maria Zimina-Poirot McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown & Jared Kaplan. 2022. Constitutional AI: Harmlessness from AI feedback. https://arxiv.org/abs/2212.08073. Denning, Peter J. & John Arquilla. 2022. The context problem in artificial intelligence. Communications of the ACM 65(12). 18–21. DOI: 10.1145/3567605. Doğru, Gökhan & Joss Moorkens. 2024. Data augmentation with translation memories for desktop machine translation fine-tuning in 3 language pairs. Journal of Specialised Translation (41). 149–178. DOI: 10.26034/cm.jostrans. 2024.4716. Dong, Qingxiu, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, Baobao Chang, Xu Sun, Lei Li & Zhifang Sui. 2024. A survey on in-context learning. https://arxiv.org/abs/2301.00234. Douglas, Michael R. 2023. Large language models. https://arxiv.org/abs/2307. 05782. EMT. 2022. European master’s in translation competence framework 2022. Tech. rep. Brussels: European Commission. 1–12. https://commission.europa.eu/ system/files/2022-11/emt_competence_fwk_2022_en.pdf. Fabricius, Anne. 2022. What’s in a sociolinguistic norm? The case of change in prevocalic /r/ in received pronunciation. In Janus Mortensen & Kamilla Kraft (eds.), Norms and the study of language in social life, 157–184. Berlin, Germany: De Gruyter. DOI: 10.1515/9781501511882-007. Farrell, Michael. 2023. Preliminary evaluation of ChatGPT as a machine translation engine and as an automatic post-editor of raw machine translation output from other machine translation engines. In Proceedings of the International Conference on Human-informed Translation and Interpreting Technology (HiTIT), 108–113. DOI: 10.26615/issn.2683-0078.2023_009. Garcia, Ignacio. 2023. Computer-aided translation systems. In Chan Sin-wai (ed.), Routledge encyclopedia of translation technology, 2nd edn., 76–94. London, UK: Routledge. DOI: 10.4324/9781003168348-4. Gledhill, Christopher & Natalie Kübler. 2016. What can linguistic approaches bring to English for specific purposes? ASp (69). 65–95. DOI: 10.4000/asp.4804. 162 8 Computer-assisted LM in teaching human-centred AT Gledhill, Christopher & Mojca Pecman. 2018. On alternating pre-modified and post-modified nominals such as aspirin synthesis vs. synthesis of aspirin: Rhetorical and cognitive packing in English science writing. Fachsprache 40(12). 24–46. DOI: 10.24989/fs.v40i1-2.1601. Gledhill, Christopher & Maria Zimina-Poirot. 2023. Human-machine interaction: How to integrate plain language rules in the revision cycles of neural machine translation output. De Europa: European and Global Studies Journal Special 2022(1). Issue). 149. https://u-paris.hal.science/hal-04050523. Hao, Yu & Anthony Pym. 2023. Where do translation students go? A study of the employment and mobility of master graduates. The Interpreter and Translator Trainer 17(2). 211–229. DOI: 10.1080/1750399X.2022.2084595. He, Zhiwei, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi & Xing Wang. 2023. Exploring human-like translation strategy with large language models. https://arxiv.org/abs/2305.04118. Hu, Bin, Chenyang Zhao, Pu Zhang, Zihao Zhou, Yuanhang Yang, Zenglin Xu & Bin Liu. 2024. Enabling intelligent interactions between an agent and an LLM: A reinforcement learning approach. https://arxiv.org/abs/2306.03604. Kim, Bokyeong, Soojin Hwang, Sanghoon Cha, Chang Hyun Park, Jongse Park & Jaehyuk Huh. 2022. Supporting dynamic translation granularity for hybrid memory systems. In 2022 IEEE 40th International Conference on Computer Design (ICCD), 25–32. DOI: 10.1109/ICCD56317.2022.00015. Martínez, Gonzalo, Lauren Watson, Pedro Reviriego, José Alberto Hernández, Marc Juarez & Rik Sarkar. 2023. Towards understanding the interplay of generative artificial intelligence and the internet. https://arxiv.org/abs/2306.06130. Mitchell-Schuitevoerder, Rosemary. 2020. A project-based approach to translation technology. London, UK: Routledge. Mosbach, Marius, Tiago Pimentel, Shauli Ravfogel, Dietrich Klakow & Yanai Elazar. 2023. Few-shot fine-tuning vs. in-context learning: A fair comparison and evaluation. https://arxiv.org/abs/2305.16938. Moslem, Yasmin, Rejwanul Haque, John D. Kelleher & Andy Way. 2023. Adaptive machine translation with large language models. In Mary Nurminen, Judith Brenner, Maarit Koponen, Sirkku Latomaa, Mikhail Mikhailov, Frederike Schierl, Tharindu Ranasinghe, Eva Vanmassenhove, Sergi Alvarez Vidal, Nora Aranberri, Mara Nunziatini, Carla Parra Escartín, Mikel Forcada, Maja Popovic, Carolina Scarton & Helena Moniz (eds.), Proceedings of the 24th Annual Conference of the European Association for Machine Translation, 227– 237. Tampere, Finland: European Association for Machine Translation. https: //aclanthology.org/2023.eamt-1.22/. 163 Maria Zimina-Poirot O’Brien, Sharon. 2024. Human-centered augmented translation: Against antagonistic dualisms. Perspectives 32(3). 391–406. DOI: 10.1080/0907676X.2023. 2247423. Pleyer, Michael & James Winters. 2015. Integrating cognitive linguistics and language evolution research. Theoria et Historia Scientiarum 11. 19. DOI: 10.12775/ ths-2014-002. Ramírez-Polo, Laura & Chelo Vargas-Sierra. 2023. Translation technology and ethical competence: An analysis and proposal for translators’ training. Languages 8(2). 93. DOI: 10.3390/languages8020093. Raus, Rachele, Alida Maria Silletti, Silvia Domenica Zollo & John Humbley (eds.). 2023. Multilinguisme et variétés linguistique en europe à l’aune de l’intelligence artificielle. Milan, Italy: Ledizioni. https://www.collane.unito.it/oa/items/ show/132. Raza, Shaina, Oluwanifemi Bamgbose, Shardul Ghuge, Fatemeh Tavakol, Deepak John Reji & Syed Raza Bashir. 2025. Developing safe and responsible large language model: can we balance bias reduction and language understanding in large language models? https://arxiv.org/abs/2404.01399. Schulz, Wolfgang & Christian Ollig. 2023. Teaching norms to large language models: The next frontier of hybrid governance. Alexander von Humboldt Institute for Internet & Society. DOI: 10.5281/ZENODO.7957165. Shichkina, Yulia & Kirill Krinkin. 2022. Principles of building personalized intelligent human assistants. In 2022 Fourth International Conference Neurotechnologies and Neurointerfaces (CNN), 148–151. DOI: 10.1109/CNN56452.2022.9912559. Sin-wai, Chan. 2023. Routledge encyclopedia of translation technology. 2nd edn. London, UK: Routledge. DOI: 10.4324/9781003168348. Sinner, Carsten. 2020. Linguistic norm in sociolinguistics. In Franz Lebsanft & Felix Tacke (eds.), Manual of standardization in the Romance languages, 145– 164. Berlin, Germany: De Gruyter. DOI: 10.1515/9783110458084-005. Slimi, Zouhaier & Beatriz Villarejo Carballido. 2023. Navigating the ethical challenges of artificial intelligence in higher education: An analysis of seven global AI ethics policies. TEM Journal. 590–602. DOI: 10.18421/TEM122-02. Tavares, Célia, Laura Tallone, Luciana Oliveira & Sandra Ribeiro. 2023. The challenges of teaching and assessing technical translation in an era of neural machine translation. Education Sciences 13(6). 541. DOI: 10.3390/educsci13060541. Uchida, Satoru. 2024. Using early LLMs for corpus linguistics: Examining ChatGPT’s potential and limitations. Applied Corpus Linguistics 4(1). 100089. DOI: 10.1016/j.acorp.2024.100089. Van Dijk, Teun A. 2006. Discourse, context and cognition. Discourse Studies 8(1). 159–177. DOI: 10.1177/1461445606059565. 164 8 Computer-assisted LM in teaching human-centred AT Yang, Jian, Hongcheng Guo, Yuwei Yin, Jiaqi Bai, Bing Wang, Jiaheng Liu, Xinnian Liang, Linzheng Cahi, Liqun Yang & Zhoujun Li. 2024. M3p: towards multimodal multilingual translation with multimodal prompt.https://arxiv.org/ abs/2403.17556. Zhu, Yilun, Joel Ruben Antony Moniz, Shruti Bhargava, Jiarui Lu, Dhivya Piraviperumal, Site Li, Yuan Zhang, Hong Yu & Bo-Hsiang Tseng. 2024. Can large language models understand context? https://arxiv.org/abs/2402.00858. Zhuang, Yan, Qi Liu, Yuting Ning, Weizhe Huang, Zachary A. Pardos, Patrick C. Kyllonen, Jiyun Zu, Qingyang Mao, Rui Lv, Zhenya Huang, Guanhao Zhao, Zheng Zhang, Shijin Wang & Enhong Chen. 2024. From static benchmarks to adaptive testing: psychometrics in AI evaluation. https://arxiv.org/abs/2306. 10512. Zimina, Maria. 2015. Interactions dans l’espace textuel multilingue. Meta 60(2). 379. DOI: 10.7202/1032930ar. Zimina, Maria. 2019. Vers une mémoire de traduction dynamique et multidimensionnelle: Traduction et technologie, regards croisés sur de nouvelles pratiques. Des mots aux actes (8). 221–236. DOI: 10.15122/ISBN.978-2-406-09779-2. Zimina, Maria & Serge Fleury. 2016. Perspectives de l’architecture trame/cadre pour les alignements multilingues. Nouvelles perspectives en sciences sociales 11(1). 325–353. DOI: 10.7202/1035940ar. Zimina-Poirot, Maria & Christopher Gledhill. 2023. The evolutive nature of translation and teaching workflow: Textometric analysis of multiple revision cycles in a simulated website translation project. mediAzioni 39. A116–A146. DOI: 10.6092/issn.1974-4382/18790. 165